Tampa BSides - Chef's Tour of Microsoft Security Adoption Framework (SAF)
Mclennan luce charlesluce
1. Journal of Experimental Psychology: Copyright 2003 by the American Psychological Association, Inc.
Learning, Memory, and Cognition 0278-7393/03/$12.00 DOI: 10.1037/0278-7393.29.4.539
2003, Vol. 29, No. 4, 539 –553
Representation of Lexical Form
Conor T. McLennan, Paul A. Luce, and Jan Charles-Luce
University at Buffalo, The State University of New York
The authors attempted to determine whether surface representations of spoken words are mapped onto
underlying, abstract representations. In particular, they tested the hypothesis that flaps—neutralized
allophones of intervocalic /t/s and /d/s—are mapped onto their underlying phonemic counterparts. In 6
repetition priming experiments, participants responded to stimuli in 2 blocks of trials. Stimuli in the 1st
block served as primes and those in the 2nd as targets. Primes and targets consisted of English words
containing intervocalic /t/s and /d/s that, when produced casually, were flapped. In all 6 experiments,
reaction times to target items were measured as a function of prime type. The results provide evidence
for both surface and underlying form-based representations.
Information-processing theories have typically characterized tations are accessed directly from binary phonetic features. Simi-
spoken word perception as being composed of a series of linguistic larly, Klatt’s (1989) LAFS (Lexical Access from Spectra) model
stages of analysis, with form-based (or sound-based) representa- proposes that only context-sensitive spectra are computed in the
tions becoming successively more abstract at each stage of pro- process of mapping waveform onto word. In short, although both
cessing. Studdert-Kennedy (1974) provided one of the earliest mediated and direct access theories assume that sensory informa-
explicit articulations of this kind of mediated lexical access model, tion is initially recoded in some manner, they differ as to whether
which itself drew inspiration from linguistic theory (see Bloom- additional levels of representation intervene between sensory re-
field, 1933; Chomsky & Halle, 1968; Harris, 1955; Kenstowicz & coding and lexical representation.
Kisseberth, 1979; see also Jusczyk & Luce, 2002, for a discus- Examples of direct and mediated models are illustrated in Fig-
sion). More contemporary examples of mediated access can be ure 1. According to extreme direct access models, auditory repre-
found in computational models of spoken word recognition such as sentations based on initial sensory recoding are mapped directly
TRACE (McClelland & Elman, 1986), Shortlist (Norris, 1994), onto form-based representations, which are then used to contact
and PARSYN (Luce, Goldinger, Auer, & Vitevitch, 2000). lemmas. In contrast, mediated access models posit some form of
Recently, these mediated access models have been challenged intermediate representations between initial recoding and lexical
by proposals that eschew the postulation of intermediate represen- representation, illustrated in Figure 1 as allophones, phonemes,
tations. According to direct access models, after the initial recod- and syllables.
ing of sensory data, information is mapped directly onto form- Evidence in support of direct access models comes from a series
based lexical representations. For example, Stevens’s lexical of experiments reported by Marslen-Wilson and Warren (1994).
access from features (LAFF) model (see Klatt, 1989) and Marslen- Building on earlier work by Whalen (1984, 1991) and Streeter and
Wilson and Warren’s (1994; see also Lahiri & Marslen-Wilson, Nigro (1979), Marslen-Wilson and Warren generated a set of
1991) direct access featural model propose that lexical represen- cross-spliced words and nonwords, creating subcategorical mis-
matches. For example, the initial consonant and vowel of the word
jog were spliced onto the final consonant of the word job, resulting
in a mismatch between the information in the vowel (which is
Conor T. McLennan and Paul A. Luce, Department of Psychology and consistent with a final /g/) and the spliced final consonant /b/. Both
Center for Cognitive Science, University at Buffalo, The State University
mediated and direct access theories predict processing costs when
of New York; Jan Charles-Luce, Department of Communicative Disorders
and Sciences and Center for Cognitive Science, University at Buffalo, The
words are cross spliced with other words because of conflicting
State University of New York. information at the lexical level. In the jog/job example, although
Portions of this work were presented at the 14th International Congress information in the vowel is consistent with the word jog, the actual
of Phonetic Sciences, August, 1999, San Francisco. This research was final consonant (/b/) is consistent with the word job. Thus, both jog
supported in part by Research Grant R01 DC 0265801 from the National and job may be activated and compete for recognition.
Institute on Deafness and Other Communication Disorders, National In- Marslen-Wilson and Warren also spliced nonwords with other
stitutes of Health. nonwords. For example, the initial consonant and vowel of the
We thank Steve Goldinger, Cynthia Connine, Ken Forster, and LouAnn nonword smod were spliced onto the final consonant of the non-
Gerken for helpful discussions regarding various aspects of this project and word smob. Although both direct and mediated access theories
Theofanis Pantazis and Melissa Pfennig for their help preparing stimuli and
predict conflicts when two words are cross spliced, only mediated
running participants.
Correspondence concerning this article should be addressed to Conor T.
access theories predict processing costs when nonwords are cross
McLennan or Paul A. Luce, Language Perception Laboratory, 245 Park spliced with other nonwords. This prediction is based on the
Hall, Department of Psychology, University at Buffalo, The State Univer- assumption that there will be conflicting cues at a sublexical level.
sity of New York, Buffalo, New York 14260. E-mail: mclennan@ In the smod/smob example, although information in the vowel is
buffalo.edu or luce@buffalo.edu consistent with a final /d/, the actual final consonant is /b/. Poten-
539
2. 540 MCLENNAN, LUCE, AND CHARLES-LUCE
Figure 1. Illustration of mediated and direct access theories.
tial conflicts may arise because some information in the input is Recently, McQueen, Norris, and Cutler (1999) challenged
consistent with the sublexical unit /d/, whereas other information is Marslen-Wilson and Warren’s finding. They found that the crucial
consistent with the sublexical unit /b/. Direct access theories pre- distinction between words cross spliced with other words and
dict no processing cost when nonwords are cross spliced with other nonwords cross spliced with other nonwords could be made to
nonwords because, according to these theories, no intermediate come and go as a function of task demands. Moreover, they found
representations exist and therefore no representations are in that models with a phonemic level of representation could simulate
conflict. the data pattern obtained by Marslen-Wilson and Warren, thus
Marslen-Wilson and Warren’s results supported the predictions calling into question the claim that mediated models should always
of direct access theories of spoken word recognition: They ob- show effects of conflicting information at a sublexical level. None-
served processing costs only when mismatching coarticulatory theless, a lack of positive evidence for sublexical representations
information involved words. The critical finding was that non- persists. As a result, the debate between mediated and direct access
words cross spliced with other nonwords failed to exhibit process- theories remains unresolved.
ing costs associated with subcategorical mismatch. Marslen- We examined the status of intermediate representations in more
Wilson and Warren concluded that the failure to find effects of detail by exploring the perceptual consequences of allophonic
subcategorical mismatch for nonwords is due to the absence of variation. More specifically, we examined flapping in American
intermediate representations. English. A flap ( ) is a neutralized version and allophone of
3. REPRESENTATION OF LEXICAL FORM 541
intervocalic /t/ and /d/. In casual American English, when a /t/ or if flapped words (e.g., activate underlying phonemic rep-
a /d/ is produced between two vowels, as in greater or Adam, it is resentations, they should prime—and be primed by— carefully
often realized as a flap, a segment that is neither exactly a /t/ nor articulated words (e.g., , ). On the other hand, if
exactly a /d/ (see Patterson & Connine, 2001). We attempted to flapped words are not mapped onto underlying phonemic repre-
determine whether flaps map onto their underlying, abstract pho- sentations, they should fail to prime their carefully articulated
nemic counterparts, /t/ and /d/. Mediated access theories predict counterparts. This latter outcome would parallel the results of
that allophonic variation occurring on the surface should map onto many previous studies (e.g., Church & Schacter, 1994; Goldinger,
more abstract, underlying phonological representations (see, e.g., 1996) in which changes in surface properties of words caused a
Pisoni & Luce, 1987). However, according to direct access theo- marked attenuation of long-term priming (dubbed specificity).
ries, allophonic variation occurring on the surface should map
directly onto lexical representations. Therefore, examining the Experiment 1
perceptual consequences of allophonic variation may help to dis-
tinguish between these competing theories. Method
The current study examined the representational status of flaps
Participants. Twenty-four participants were recruited from the Uni-
in memory using a repetition priming paradigm. In this paradigm, versity at Buffalo community. They were paid $5 or received partial credit
participants are presented with a block of spoken words to which for a course requirement. Participants were right-handed native speakers of
they must respond (the study phase). After this initial exposure, American English, with no reported history of speech or hearing disorders.
participants are presented with another block of words (the test Materials. The stimuli consisted of bisyllabic spoken words containing
phase). In the second block, some of the words from the first block alveolar and non-alveolar medial consonants. The alveolar stimuli con-
are repeated. Typically, repeated words are responded to more sisted of 12 sets of spoken words. Each set contained three stimuli: a
quickly and accurately than new words (Church & Schacter, 1994; minimal pair of carefully produced words that differed only on the voicing
Goldinger, 1996; Kempley & Morton, 1982; Luce & Lyons, 1998). of the medial alveolar stop (e.g., and and a casually
This repetition priming effect presumably arises because repeated produced flapped version of the minimal pair (e.g., ). The non-
alveolar stimuli consisted of 12 words containing a medial [b], [p], [g], or
activation of form-based representation facilitates processing.
[k] (e.g., bacon). The non-alveolar stimuli were also casually and carefully
Repetition priming can be used to determine whether two nom- produced. Casually and carefully produced stimuli differed primarily in
inally different stimuli activate the same mental representation. In speed of articulation. In addition, casual alveolar stimuli were produced
particular, the priming paradigm may be used to determine with a flap. Note that flapped stimuli may be ambiguous (i.e., may
whether flapped segments are mapped onto underlying intermedi- refer to atom or Adam), whereas casually produced non-alveolar stimuli are
ate form-based representations of /t/s, /d/s, or both, or whether not. A complete list of the stimuli used in all experiments is presented in
flaps are represented veridically as they appear in casual speech the Appendix.
as . Specifically, is there a recoding of the surface allophonic The final 12 sets of alveolar stimuli were chosen from 24 sets of
representation, , to the underlying phonological representation, carefully and casually articulated words containing intervocalic /t/s and
/t/ or /d/, as predicted by mediated access theories of spoken word /d/s. As a means of ensuring that the casually articulated alveolar stimuli
contained fully ambiguous flapped segments (and not clear /t/s or /d/s),
recognition?
the 72 stimuli composing the 24 sets in the original master list were
In the present experiments, two blocks of stimuli containing randomized and presented to 10 listeners in a forced-choice (/t/–/d/) iden-
carefully and casually articulated versions of words (and non- tification task. Twelve flapped stimuli were then chosen that were not
words) were presented. Casually articulated (hypoarticulated) identified consistently as containing a /t/ or /d/ by more than 6 listeners. On
words are produced in a relaxed manner, whereas carefully artic- average, half of the participants identified the 12 flapped stimuli as con-
ulated words are more clearly articulated. Intervocalic /t/s and /d/s taining /d/ and half as /t/. Thus, flapped words were perceived to be
are flapped in casually articulated words but not in carefully ambiguous. For the carefully articulated stimuli, 9 or more listeners iden-
articulated words. We hypothesize that priming of casually artic- tified the stimuli as containing the intended segment.
ulated stimuli by carefully articulated stimuli (or vice versa) indi- The /t/ and /d/ members of the stimulus pairs were matched on average
cates the presence of a mediating underlying representation in log frequency of occurrence (Kucera & Francis, 1967). The mean log
ˇ
frequencies for /t/ and /d/ words were .53 and .30, respectively. This
memory. We refer to any significant attenuation in priming for
difference was not significant, t(11) ϭ 1.67, p ϭ .12. The mean durations
stimuli that mismatch in articulation style as evidence for speci- for /d/ and /t/ carefully articulated words were 529 ms and 515 ms,
ficity. The presence of specificity effects indicates the absence of respectively. This difference was not significant, t(11) ϭ 0.53, p ϭ .61.
intermediate representations, consistent with direct access theories. The mean duration for the flapped stimuli was 387 ms. The difference in
Conversely, lack of specificity effects indicates the presence of duration between the casual (flapped) and careful stimuli reflects articula-
intermediate representations, consistent with mediated access tion style; no attempt was made to equate the durations of the flapped and
theories. careful stimuli.
To review, traditional information-processing theories assume The stimuli were recorded in a sound-attenuated room by a phonetically
that access to the lexicon is mediated by intervening representa- sophisticated male speaker of a midwestern dialect, low-pass filtered at 10
tions. Direct access theories assume that, after initial sensory kHz, and digitized at a sampling rate of 20 kHz with a 16-bit analog-to-
digital converter. All words were edited into individual files and stored on
registration, access to the lexicon is direct. These classes of theo-
computer disk.
ries make opposite predictions regarding the perceptual conse- Design. Two blocks of stimuli were presented. The first constituted the
quences of allophonic variation. To evaluate these theories, we primes and the second the targets. The carefully and casually produced
conducted a series of long-term repetition priming experiments in alveolar and non-alveolar stimuli served as both primes and targets. For
which flapped and carefully articulated words served as both both the primes and targets, half of the alveolar and non-alveolar stimuli
primes and targets. The basic logic of all of the experiments is that were casually articulated and half were carefully articulated. Primes
4. 542 MCLENNAN, LUCE, AND CHARLES-LUCE
matched, mismatched, or were unrelated to the targets. Matching primes uli.1 Effects are significant at the .05 level unless otherwise indi-
and targets were identical (e.g., and ). Mismatching cated. Accuracy was greater than 97% and produced no significant
primes and targets differed in articulation style only (e.g., and effects.
). The prime block consisted of 8 alveolar, 8 non-alveolar, and 8 Alveolar stimuli. RTs for the alveolar stimuli as a function of
unrelated (i.e., control) stimuli. The target block consisted of 12 alveolar
prime and target type are plotted in the upper left panel of Figure 2.
stimuli and 12 non-alveolar stimuli. In the target block, 8 stimuli were
Mean RTs as a function of condition and magnitudes of specificity
matching, 8 were mismatching, and 8 were control.
Orthogonal combination of the three levels of prime type (match, mis- and priming for all six experiments are shown in Table 2. Mag-
match, and control) and two levels of target type (casual and careful) nitude of specificity is indicated by the difference in RT between
resulted in six conditions, shown in Table 1. Across participants, each the matching and mismatching conditions. Magnitude of priming
careful and casual item was present in every possible condition. However, is indicated by the difference in RT between the matching and
no single participant heard more than one version of a given word within control conditions.
a block. For example, if a participant heard the word in one of the Casually articulated (i.e., flapped) items were responded to more
blocks, he or she did not hear , , or again in the quickly than carefully articulated items, F1(1, 22) ϭ 56.08,
same block. MSE ϭ 5,934.54, F2(1, 11) ϭ 24.53, MSE ϭ 7,004.35, presumably
Procedure. Participants were tested individually in a quiet room and
because of the differences in duration. There was also a significant
were not told at the beginning of the experiment that there would be two
effect of prime type, F1(2, 44) ϭ 3.72, MSE ϭ 6,402.81, F2(2,
blocks of trials. Participants performed a single-word shadowing task in
which they attempted to repeat (or shadow) the stimulus word as quickly 22) ϭ 3.26, MSE ϭ 5,532.03. Prime type and target type did not
and accurately as possible. In both the prime and target blocks, the stimuli interact, F1 and F2 Ͻ 1.
were presented binaurally over headphones. The headphones had an at- Planned comparisons based on the main effect of prime type
tached microphone that was placed approximately 1 in. (2.5 cm) from the revealed significant differences between match and control condi-
participant’s lips. A Centris 650 computer controlled stimulus presentation tions and between mismatch and control conditions, F1(1,
and recorded shadowing times. Stimulus presentation within each block 44) ϭ 6.34, F2(1, 22) ϭ 5.47 and F1(1, 44) ϭ 4.70, F2(1,
was random for each participant. 22) ϭ 4.23, respectively. There was no difference between match
A given trial proceeded as follows. A light at the top of the response box and mismatch conditions, F1 and F2 Ͻ 1.
was illuminated to indicate the beginning of the trial. The participant was
Both matching and mismatching prime types produced signifi-
then presented with a stimulus word binaurally over the headphones. The
cant facilitative effects on shadowing times. Furthermore, match-
participant was instructed to shadow the stimulus word as quickly and
accurately as possible. Reaction times (RTs) were measured from onset of ing primes facilitated target shadowing as much as mismatching
the presentation of the stimulus word to onset of the participant’s shad- primes. These results are consistent with theories that posit under-
owing response. After the participant responded, the next trial was initi- lying intermediate representations.
ated. If the maximum RT (5 s) expired, the computer automatically Non-alveolar stimuli. RTs for the non-alveolar stimuli as a
recorded an incorrect response and presented the next trial. function of prime and target type are plotted in the upper right
panel of Figure 2. Magnitudes of specificity and priming are
Results shown in Table 2. Casually articulated items were responded to
more quickly than carefully articulated items, F1(1, 22) ϭ 24.32,
RTs less than 200 ms or greater than 2,000 ms were replaced MSE ϭ 9,764.20, F2(1, 11) ϭ 11.27, MSE ϭ 10,663.38. There was
with the appropriate condition mean. Less than 1% of the RTs also a significant effect of prime type, F1(2, 44) ϭ 3.78,
were replaced. Any participant whose overall mean RT fell two MSE ϭ 7,246.919, F2(2, 22) ϭ 4.83, MSE ϭ 3,696.95. Prime type
standard deviations beyond the grand mean was excluded, result- and target type did not interact, F1(2, 44) ϭ 1.13, MSE ϭ 8,858.18,
ing in the elimination of 1 participant. F2 Ͻ 1.
Prime Type (match, mismatch, or control) ϫ Target Type (care-
ful or casual) participant (F1) and item (F2) analyses of variance
(ANOVAs) were performed on RTs for correct responses and 1
In this and all subsequent experiments, Prime Type ϫ Target Type ϫ
percentages correct for the alveolar and non-alveolar target stim- Voicing (/t/ vs. /d/) analyses were first performed. In no instance did
voicing enter into any significant interactions. Thus, in all analyses, we
collapsed across /t/ and /d/ stimuli. In addition, when appropriate (i.e.,
Experiments 1, 2, 3, and 4), analyses were performed that compared
Table 1 relative changes in RTs from prime to target block. These analyses were
Experimental Conditions and Examples consistent with all analyses on the targets alone in this and subsequent
experiments. Finally, for a number of reasons, item analyses may not be
Example appropriate for the current experiments. First, the stimuli used exhaust the
(small) universe of items that meet our specific criteria, making the need
Condition Block 1: prime Block 2: target
for generalization beyond the present set of stimuli unnecessary. Second,
Match the stimuli are matched on all variables known to affect the dependent
Careful prime 3 careful target ætəm ætəm variables under scrutiny, thus calling into question the suitability of per-
Casual prime 3 casual target æɾəm æɾəm forming traditional ANOVAs with items as random factors (see Raaijmak-
Mismatch ers, Schrijnemakers, & Gremmen, 1999). Finally, the low number of items
Casual prime 3 careful target æɾəm ætəm meeting our stringent criteria unavoidably reduces the statistical power of
Careful prime 3 casual target ætəm æɾəm our tests. Despite these caveats, we nonetheless report item analyses, more
Control because of convention than because of their appropriateness. Readers
Unrelated prime 3 careful target pep ætəm should bear in mind these caveats in interpreting the significance levels of
Unrelated prime 3 casual target pep æɾəm
all item tests reported for the current studies.
5. REPRESENTATION OF LEXICAL FORM 543
Figure 2. Top: Mean reaction times (RTs) for the alveolar (left) and non-alveolar (right) stimuli in Experi-
ment 1. Bottom: Mean RTs for the alveolar (left) and non-alveolar (right) stimuli in Experiment 2.
Planned comparisons based on the main effect of prime type times for the non-alveolar stimuli demonstrated that words match-
revealed significant differences between match and control condi- ing in articulation style were more effective primes for casually
tions and between match and mismatch conditions, F 1 (1, and carefully articulated non-alveolar targets than mismatching
44) ϭ 5.25, F2(1, 22) ϭ 8.22 and F1(1, 44) ϭ 6.08, F2(1, words.
22) ϭ 6.10, respectively. There was no difference between mis- These results suggest that underlying intermediate representa-
match and control conditions, F1 and F2 Ͻ 1. tions are activated during processing of phonologically ambiguous
For the non-alveolar stimuli, which did not contain flaps, facil- flapped stimuli. However, in the absence of ambiguity, surface
itative priming was observed only when production style (careful representations appear to suffice, as evidenced by the non-alveolar
and casual) matched. These results are consistent with theories that stimuli. Thus, the present data provide evidence for the existence
posit distinct surface representations but contrast with the results
of both surface and underlying lexical representations in memory.
obtained for the alveolar stimuli.
As a result, these findings join a growing body of evidence in
support of lexical representations that preserve surface information
Discussion (e.g., Church & Schacter, 1994; Goldinger, 1996) while calling
Experiment 1 revealed two notable findings. First, the shadow- into question a purely instance- or exemplar-based model of the
ing times for the alveolar stimuli demonstrated that casually and mental lexicon (e.g., Goldinger, 1998).
carefully articulated words are equally effective primes for both Another possible explanation for the lack of specificity observed
casually and carefully articulated targets. Second, the shadowing for the alveolar stimuli is that lemmas (i.e., semantic–syntactic
6. 544 MCLENNAN, LUCE, AND CHARLES-LUCE
Table 2
Reaction Times, Standard Errors, and Magnitudes of Specificity and Priming for
Experiments 1– 6
Reaction time (ms)
Match Mismatch Control
Experiment Stimuli M SE M SE M SE MOS MOP
1. Shadowing Alveolar 844 18 850 18 886 22 Ϫ6 Ϫ42*†
Non-alveolar 859 21 903 20 900 19 Ϫ44*† Ϫ41*†
2. Shadowing Alveolar 817 20 829 20 871 22 Ϫ12 Ϫ54*††
Non-alveolar 822 21 877 20 879 21 Ϫ55*† Ϫ57*†
3. EDLD Alveolar 819 24 934 34 914 26 Ϫ115*† Ϫ95*†
4. HDLD Alveolar 953 25 997 29 1071 33 Ϫ44 Ϫ118*†
5. EDLD–shadowing Alveolar 874 15 861 15 904 17 13 Ϫ30*††
6. Shadowing–EDLD Alveolar 873 17 870 21 929 24 3 Ϫ56*††
Note. MOS ϭ magnitude of specificity (match Ϫ mismatch); MOP ϭ magnitude of priming (match Ϫ
control). EDLD ϭ easy-discrimination lexical decision; HDLD ϭ hard-discrimination lexical decision.
* p Ͻ .05, by participants. † p Ͻ .05, by items. †† .07 Ͼ p Ͼ .05, by items.
representations), and not intermediate form-based representations, tions also dominate processing for carefully articulated alveolar
may have mediated the priming effect. For example, the ambigu- stimuli may be premature. Given the potentially important theo-
ous flapped stimulus may have activated the lemmas for retical implications of these findings, we attempted to replicate
both atom and Adam, which in turn may have facilitated process- Experiment 1. Moreover, combining the data from Experiments 1
ing of the lemmas as targets. If this is the case, there is no need to and 2 should increase the power of the statistical tests for detecting
posit activation of underlying intermediate form-based represen- what may be a weak effect of specificity.
tations corresponding to /t/ or /d/: The facilitative effect of prime
on target may have emanated exclusively from the semantic level. Experiment 2
The data for the non-alveolar stimuli contradict this hypothesis.
Clearly, if the long-term repetition priming effect is lemma based, Method
mismatches in articulation style should have no effect on the
Participants. A different group of 24 participants were recruited from
magnitude of facilitative priming. However, the results for the
the University at Buffalo community. They were paid $5 or received partial
non-alveolar stimuli revealed facilitation only when stimuli credit for a course requirement. Participants met the same criteria as those
matched. For example, casually articulated bacon failed to prime in Experiment 1.
carefully articulated bacon. If the priming effect were lemma Materials and procedure. The materials and procedure were identical
based, we would have expected little or no diminution of priming to those used in Experiment 1.
as a function of differences in articulation style.
The literature is also replete with demonstrations that long-term Results
repetition priming is primarily form based. For example, long-term
repetition priming is typically modality specific (e.g., Jackson & Less than 3% of the RTs and no participants were excluded from
Morton, 1984). Were the effect lemma based, changes in modality the analyses. Accuracy was greater than 90% and produced no
should have no effect on facilitative priming. Moreover, non- significant outcomes.
words—which, by definition, have no semantic representations— Alveolar stimuli. RTs for the alveolar stimuli as a function of
show long-term repetition priming (e.g., Fisher, Hunt, Chambers, prime and target type are plotted in the lower left panel of Figure 2.
& Church, 2001; Goldinger, 1998). Finally, whether participants’ Magnitudes of specificity and priming are shown in Table 2.
attention is focused on the sound or the meanings of words does Casually articulated (i.e., flapped) items were again responded to
not appear to affect long-term repetition priming (e.g., Church & more quickly than carefully articulated items, F1(1, 23) ϭ 47.27,
Schacter, 1994). In short, given previous findings—as well as our MSE ϭ 5,053.58, F2(1, 11) ϭ 12.43, MSE ϭ 10,578.86. And
own results for the non-alveolar stimuli—we can be confident that again, we obtained a main effect of prime type, F1(2, 46) ϭ 3.31,
the locus of the effect is at the form level. MSE ϭ 11,964.97, F2(2, 22) ϭ 2.14, MSE ϭ 8,412.95, p ϭ .14.
We observed a small numerical trend toward specificity among Most important, prime type and target type did not interact, F1 and
the carefully articulated alveolar stimuli (match: 890 ms; mis- F2 Ͻ 1.
match: 905 ms; see upper left panel of Figure 2). Therefore, it is Planned comparisons revealed a significant difference between
possible that the lack of interaction between prime and target type the match and control conditions, F1(1, 46) ϭ 6.02, F2(1,
is due to a lack of power. In other words, low power might have 22) ϭ 3.96, p ϭ .059; the difference between the mismatch and
been at least partially responsible for the lack of a specificity effect control conditions was marginally significant by participants but
among the carefully articulated alveolar items in Experiment 1. If not by items, F1(1, 46) ϭ 3.59, p ϭ .064, F2(1, 22) ϭ 2.19, p ϭ
this is the case, ruling out the possibility that surface representa- .153. However, the difference between the match and mismatch
7. REPRESENTATION OF LEXICAL FORM 545
conditions was not significant, F1 and F2 Ͻ 1. Aside from the specific surface representations are responsible for long-term rep-
somewhat weaker statistical outcomes, the present results replicate etition priming.
those obtained in Experiment 1. Note that casual articulation of the alveolar, but not the non-
Non-alveolar stimuli. RTs for the non-alveolar stimuli as a alveolar, stimuli produces phonological (and lexical) ambiguity.
function of prime and target type are plotted in the lower right For example, casual production of the word atom (an
panel of Figure 2. Magnitudes of specificity and priming are alveolar stimulus item) is ambiguous between and ;
shown in Table 2. The results for the non-alveolar stimuli also however, casual production of the word bacon (a non-alveolar
replicated Experiment 1: Casually articulated items were again stimulus item) is unambiguous. In other words, flaps map onto
responded to more quickly than carefully articulated items, F1(1, two possible underlying phonological (and lexical) representa-
23) ϭ 23.56, MSE ϭ 12,177.54, F2(1, 11) ϭ 17.69, MSE tions, whereas casual productions of non-alveolar stimuli have
ϭ 8,133.70. There was also a significant effect of prime type, F1(2, only one corresponding representation. This distinction is presum-
46) ϭ 8.89, MSE ϭ 5,675.61, F2(2, 22) ϭ 8.36, MSE ϭ 3,391.67. ably what led to the pronounced difference between the two sets of
Prime type and target type did not interact, F1 and F2 Ͻ 1. stimuli.
Planned comparisons revealed significant differences between The finding that alveolar items activate underlying representa-
match and control conditions and between match and mismatch tions, whereas non-alveolar stimuli appear to contact only highly
conditions, F1(1, 46) ϭ 13.82, F2(1, 22) ϭ 15.84 and F1(1, specific surface representations, can be accounted for within a
46) ϭ 12.84, F2(1, 22) ϭ 7.83, respectively. There was no differ- resonance framework similar to that proposed by Vitevitch and
ence between mismatch and control conditions, F1 Ͻ 1 and F2(1, Luce (1999) and based on Grossberg’s ARTPHONE model
22) ϭ 1.40. (Grossberg, Boardman, & Cohen, 1997). According to this frame-
work (illustrated in Figure 3), acoustic–phonetic input activates
Combined Analyses for Experiments 1 and 2: Alveolar chunks corresponding to sublexical and lexical representations
(only lexical representations are illustrated). (A chunk can be
Stimuli
Not surprisingly, the difference between the casually articulated
(i.e., flapped) items and the carefully articulated items was signif-
icant, F1(1, 46) ϭ 103.56, MSE ϭ 566,738.17, F2(1, 23) ϭ 35.66,
MSE ϭ 8,467.82. The combined analyses also revealed a signifi-
cant main effect of prime type, F1(2, 92) ϭ 6.89, MSE ϭ 9,065.53,
F2(2, 46) ϭ 5.37, MSE ϭ 6,676.88. Crucially, despite the in-
creased power obtained by combining the analyses from Experi-
ments 1 and 2, the interaction between prime type and target type
failed to reach significance, F1 and F2 Ͻ 1. These results suggest
that the findings of Experiments 1 and 2 are not simply due to lack
of statistical power.
Planned comparisons based on the main effect of prime type
revealed significant differences between the match and control
conditions and between the mismatch and control conditions, F1(1,
92) ϭ 12.21, F2(1, 46) ϭ 9.51 and F1(1, 92) ϭ 8.01, F2(1,
46) ϭ 6.24, respectively. The difference between the match and
mismatch conditions for the alveolar stimuli once again failed to
reach significance, F1 and F2 Ͻ 1.
Discussion
The data for both Experiments 1 and 2 revealed no statistically
significant effects of specificity (articulation style) for alveolar
stimuli, whereas differences in articulation style for non-alveolar
stimuli completely blocked facilitative priming. In both experi-
ments, we obtained evidence that a flap activates its underlying
representations. Specifically, presentation of a flapped item facil-
itated processing of items containing either /t/ or /d/, and presen-
tation of carefully articulated items containing /t/ or /d/ facilitated
processing of flapped stimuli. The priming of flaps by carefully
produced items, and vice versa, indicates that shared underlying
representations are activated during processing. On the other hand,
the pattern of results for the non-alveolar stimuli was markedly Figure 3. Illustration of the proposed resonances between input and
different: Primes facilitated their corresponding targets only when chunks. Ovals correspond to input, rectangles correspond to chunks, and
articulation style matched. In contrast to the results for the alveolar double-sided arrows correspond to resonances (i.e., percepts). For simplic-
stimuli, the data for the non-alveolar items indicate that highly ity, only lexical chunks are shown.
8. 546 MCLENNAN, LUCE, AND CHARLES-LUCE
thought of as a learned set of associated features that may vary in surface forms that preserve detail, hence producing marked effects
size from allophone to word.) These chunks then resonate with the of specificity in priming.
input, with the resonance between input and chunk constituting the Before proceeding, we should note that the adaptive resonance
percept (Grossberg, 1986). We propose that input takes the form of account also suggests why we observed a numerical (but not
specific and relatively veridical surface representations that pre- statistically significant) data pattern for the carefully produced
serve articulation style (including allophonic variation). These alveolar stimuli that is somewhat suggestive of specificity (see
surface representations resonate with chunks that correspond to Figure 2). That is, RTs for these careful target stimuli tended to be
both allophonic and more abstract phonemic representations (see somewhat slower in the mismatching than matching conditions,
Figure 3b). That is, when confronted with phonologically and consistent with some degree of specificity in long-term priming.
lexically ambiguous flapped stimuli, sublexical and lexical chunks On the basis of the current framework, attenuation in priming for
corresponding to a flapped representation, , and both underly- the carefully produced stimuli in the mismatching condition might
ing /t/ and /d/ resonate with the surface representation. Underlying be expected given that the prime is actually a restored or instan-
/t/ and /d/ chunks resonate to flapped segments because of learned tiated representation based on processing of the flapped stimulus,
associations between flaps and those lemmas that are also associ- which may serve as a less effective long-term prime.
ated with form-based representations having fully specified /t/ and According to the resonance framework, instantiation or restora-
/d/ medial stops. tion of the surface form by the underlying chunks should require
Once activated, the chunks corresponding to underlying /t/ and time. Thus, a task that taps into the recognition process before
/d/ will establish resonances with surface representations. In the restoration of the underlying form should show strong effects of
case of flapped input, no surface representation will match the specificity in long-term repetition priming, given that the under-
activated /t/ or /d/ chunks exactly. We propose that in the absence lying representations may not have had sufficient time to establish
of an exactly corresponding surface form, the activated chunk will resonance with a restored surface form. Although the single-word
itself instantiate a surface representation with which it will reso- shadowing task typically produces fairly rapid responses, we ex-
nate most strongly (see Figure 3c). We envision this process to be pect that the need to contact a representation that drives the
much like the one proposed by Grossberg and Meyers (2000, p. production response will allow—indeed encourage—the establish-
738) to account for phoneme restoration: ment of resonances between underlying and restored surface
forms. We should note that our working assumption is that under-
In phonemic restoration experiments, broadband noise may be per- lying forms will always instantiate surface forms when there is
ceived as different phonemes depending on the context. These per- phonological or lexical ambiguity. However, it should be pos-
cepts may be attributed to a process by which active list chunks use sible to devise a situation in which we tap the recognition process
their learned top-down expectations to select the noise components before the restoration of the surface form by the underlying
that are consistent with the expected formations and suppress those representations.
that are not (Grossberg, 1995, 1999). To test this hypothesis, we conducted two auditory lexical
decision experiments in which we manipulated the time required to
Thus, the ambiguous flap is analogous to a noise segment and is decide whether a spoken item is a word or nonword. In Experi-
perceived in the context of resonating list chunks that correspond ment 3, we made the word–nonword discrimination task easy by
to underlying /t/ and /d/. Indeed, we propose that activation of the including very un-wordlike nonwords (e.g., thushthudge). When
phonemic chunks by the ambiguous flap results in restoration of a presented with nonwords whose sound patterns bear relatively
surface representation not actually present in the input. As stated little resemblance to real words, participants in the lexical decision
by Grossberg and Stone (1986), “top down signal patterns . . . con- task should be able to base their decisions on overall lexical
stitute the read out of optimal templates [e.g., phonemic chunks] in activity in the system, rather than a near exhaustive analysis of the
response to ambiguous or novel bottom-up signals [e.g., stimulus itself (Luce & Pisoni, 1998; see also Coltheart, Davelaar,
flaps] . . . to form completed composite patterns that are a mixture Jonasson, & Besner, 1977). If the sound patterns of the nonwords
of actual and expected information” (p. 58). These completed are quite dissimilar to those of words, they should produce little
composite patterns serve as the basis for the long-term priming lexical activity. Thus, only a modicum of lexical activation should
effect. signal the presence of a real lexical item, thus allowing for a rapid
In the case of the non-alveolar stimuli—which showed evidence lexical decision response. In short, easy discrimination in this task
of complete specificity in long-term priming—resonances between should encourage fast processing of the word stimuli.
surface forms and underlying chunks again serve as the percept. On the other hand, difficult word–nonword discrimination (Ex-
However, given the absence of phonological and lexical ambigu- periment 4) should slow processing. If participants hear very
ity, the underlying chunks simply resonate with the surface forms wordlike nonwords (e.g., bacov, created from the word bacon),
to which they match and do not require the restoration of forms not processing of the word stimuli should require more than a super-
present in the input. Hence, the surface forms that mediate the ficial assessment of lexical activity, given that the nonwords them-
priming for the non-alveolar stimuli preserve their specific selves should strongly activate similar lexical items in memory.
characteristics. Note, however, that we expect longer RTs in both Experiments 3
To review, we propose that because of phonological and lexical and 4 than in Experiments 1 and 2 because of the additional
ambiguity, underlying representations (or chunks) activated by processing required to make a lexical decision.
flaps restore surface representations that serve as the basis for By manipulating ease of discrimination, we were able to test the
long-term repetition priming. In the absence of ambiguity (i.e., for hypothesis that instantiation of surface forms by underlying rep-
our non-alveolar stimuli), underlying representations resonate with resentations takes time. We predict that in the easy-discrimination
9. REPRESENTATION OF LEXICAL FORM 547
lexical decision task, marked effects of specificity should be ob- 44) ϭ 7.67, F2(1, 22) ϭ 8.56 and F1(1, 44) ϭ 11.30, F2(1,
served for the alveolar items, because lexical decisions in the target 22) ϭ 10.14, respectively. However, the difference between the
block should be accomplished before instantiation of a surface mismatch and control conditions was not significant, F1 and
form corresponding to the underlying representation. However, F2 Ͻ 1.
when discrimination is difficult, we predict that effects of under-
lying representations should once again be detectable in long-term Discussion
priming.
The use of the lexical decision task also allowed us to determine Matched primes produced significant facilitative effects on RTs
the degree to which the activation of underlying representations by to targets, whereas mismatched primes failed to do so: Facilitative
flapped items is dependent on the shadowing task itself. It is priming was observed only when production style (careful and
possible that the underlying representations mediating the priming casual) matched. These results are consistent with surface theories
effects in Experiments 1 and 2 are fundamentally in the service of that posit separate representations for casually and carefully artic-
speech production and have little relevance to perception. Thus, ulated stimuli but contrast with the results obtained in Experi-
replicating our results in the hard-discrimination lexical decision ments 1 and 2.
task would enable us to determine to what extent the speech As predicted, the lexical decision task with easily discriminated
production mechanism must be involved in the activation of un- words and nonwords produced specificity effects for the alveolar
derlying representations. stimuli, in contrast to Experiments 1 and 2. We hypothesize that
the easy-discrimination lexical decision task taps the recognition
Experiment 3 process before the underlying representations for phonemic /t/ and
/d/ have had time to establish resonance with a restored surface
Method form (which, according to our hypothesis, serves as the basis for
the long-term repetition priming effect). Hence, repetition effects
Participants. A different group of 24 participants were recruited from
were observed only for those stimuli matching in articulation style.
the University at Buffalo community. They were paid $5 or received partial
credit for a course requirement. Participants met the same criteria as those
To garner further evidence for this hypothesis, we conducted
in Experiment 1. another lexical decision experiment in which word–nonword dis-
Materials. The materials were the same as in Experiments 1 and 2 with crimination was made more difficult. We hypothesize that the
one exception. To create the lexical decision task, we replaced the non- additional processing required to make the more difficult lexical
alveolar stimuli used in Experiments 1 and 2 with low phonotactic prob- decision should enable underlying abstract representations suffi-
ability nonwords (e.g., thushshug). However, all of the nonwords used in cient opportunity to establish resonance with the surface form, thus
this experiment were phonotactically legal in English. attenuating the specificity effect.
Procedure. Except for the task, the procedure was the same as in
Experiments 1 and 2. Participants performed a lexical decision task in
which they were instructed to decide as quickly and accurately as possible Experiment 4
whether the item they heard was a real English word or a nonword. They
indicated their decision by pressing one of two appropriately labeled Method
buttons (word on the right and nonword on the left) on a response box Participants. A different group of 24 participants were recruited from
positioned directly in front of them. the University at Buffalo community. They were paid $5 or received partial
credit for a course requirement. Participants met the same criteria as those
Results in Experiment 1.
Materials. All materials were the same as in Experiment 3, with one
RTs less than 500 ms or greater than 2,500 ms were replaced exception: The nonwords were created from the non-alveolar stimuli used
with the appropriate condition mean.2 Less than 3% of the RTs in the first two experiments by changing the word endings (e.g., bacon 3
and 1 participant were excluded from the analyses. Accuracy was bacov), resulting in more wordlike nonwords and presumably more diffi-
greater than 87% and produced only one significant outcome. cult discrimination between words and nonwords.
Procedure. The procedure was the same as in Experiment 3.
There was a main effect of prime type, F1(2, 44) ϭ 4.07, MSE ϭ
271.74, F2(2, 22) ϭ 3.54, MSE ϭ 118.17, that was entirely driven
by low accuracy in the control condition. We report data only for Results
the words.
RTs less than 500 ms or greater than 2,500 ms were replaced
RTs as a function of prime and target type are plotted in the
with the appropriate condition mean. Less than 8% of the RTs
upper left panel of Figure 4. Magnitudes of specificity and priming
were excluded from the analyses. In addition, 1 participant was
are shown in Table 2. Casually articulated (i.e., flapped) items
excluded. Accuracy was greater than 81% and produced no sig-
were responded to more quickly than carefully articulated items,
nificant outcomes. As in Experiment 3, we report data only for the
F1(1, 22) ϭ 6.09, MSE ϭ 36,670.94, F2(1, 11) ϭ 10.49,
words.
MSE ϭ 10,325.40. There was also a main effect of prime type,
RTs as a function of prime and target type are plotted in the
F1(2, 44) ϭ 6.44, MSE ϭ 26,620.07, F2(2, 22) ϭ 6.25,
upper right panel of Figure 4. Magnitudes of specificity and
MSE ϭ 19,926.17. Prime type and target type did not interact, F1
and F2 Ͻ 1.
Planned comparisons based on the main effect of prime type 2
Different upper and lower cutoffs were employed for the two types of
revealed significant differences between the match and control tasks (shadowing and lexical decision) because of the overall longer RTs in
conditions and between the match and mismatch conditions, F1(1, the lexical decision task.
10. 548 MCLENNAN, LUCE, AND CHARLES-LUCE
Figure 4. Top: Mean reaction times (RTs) for the stimuli in Experiments 3 (left) and 4 (right). Bottom: Mean
RTs for the stimuli in Experiments 5 (left) and 6 (right).
priming are shown in Table 2. Casually articulated (i.e., flapped) Discussion
items were responded to more quickly than carefully articulated
items, F1(1, 22) ϭ 8.40, MSE ϭ 32,019.86, F2(1, 11) ϭ 8.71, As predicted, increasing the difficulty of word–nonword dis-
MSE ϭ 21,326.97. There was also a main effect of prime (signif- crimination in the lexical decision task attenuates specificity ef-
icant by participants and marginal by items), F1(2, 44) ϭ 6.63, fects.3 We propose that more difficult lexical discrimination forces
MSE ϭ 24,764.27, F2(2, 22) ϭ 2.88, MSE ϭ 16,925.01, p ϭ .077. a more exhaustive analysis of the stimulus, resulting in increased
Prime type and target type did not interact, F1 and F2 Ͻ 1. opportunities for underlying representations to establish resonance
Planned comparisons based on the main effect of prime type with restored surface representations. Despite a numerical trend
revealed significant differences between the match and control
conditions and between the mismatch and control conditions (al- 3
though the latter effect was marginal by items), F1(1, 44) ϭ 13.01, If our manipulation was successful in increasing the difficulty of
word–nonword discrimination, RTs to target items should be significantly
F2(1, 22) ϭ 5.45 and F1(1, 44) ϭ 5.03, F2(1, 22) ϭ 2.72, p ϭ .113,
longer in Experiment 4 (difficult discrimination) than in Experiment 3
respectively. However, the difference between the match and mis- (easy discrimination). To confirm the effectiveness of manipulating the
match conditions was not significant, F1(1, 44) ϭ 1.86, F2 Ͻ 1. wordlikeness of the nonwords, we performed an ANOVA on mean RTs to
Matched and mismatched primes produced significant facilitative target items in Experiments 3 (M ϭ 888.86 ms) and 4 (M ϭ 1,007.10 ms).
effects on target RTs. These results replicated those for the alve- The main effect of experiment was significant, indicating that our manip-
olar items in Experiments 1 and 2. ulation was indeed successful.
11. REPRESENTATION OF LEXICAL FORM 549
toward specificity in Experiment 4, the overall pattern of results credit for a course requirement. Participants met the same criteria as those
supports the hypothesis that increasing the depth of processing in Experiment 1.
should result in stronger resonance of underlying representations Materials. The materials in Block 1 were the same as those in Exper-
with surface forms (see also Goldinger, 1996).4 iment 3, and the materials in Block 2 were the same as those in Experi-
ments 1 and 2.
The evidence presented thus far suggests that underlying pho-
Procedure. The procedure was the same as in the previous experiments
nemic representations are contacted during recognition, but only
with one exception. In the first block participants performed a lexical
under circumstances in which resonances between underlying and decision task, and in the second block participants performed a shadowing
surface forms are encouraged to develop. In particular, we see task.
evidence of underlying representations in long-term repetition
priming when a certain degree of depth of processing is required, Results
either by having to generate a production response in the shadow-
ing task or by having to make a difficult word–nonword discrim- RTs less than 200 ms or greater than 2,000 ms were replaced
ination in the lexical decision task.5 with the appropriate condition mean. Less than 2% of the RTs
A subtle, but potentially important, alternative hypothesis re- were excluded from the analyses. In addition, 1 participant was
garding our depth of processing account deserves consideration. excluded. Accuracy was greater than 97% and produced no sig-
Perhaps underlying forms always establish resonances with sur- nificant outcomes.
face forms, regardless of the circumstances. However, the point at RTs as a function of prime and target type are plotted in the
which the response taps the perceptual process may be the deter- lower left panel of Figure 4. Magnitudes of specificity and priming
mining factor in whether repetition priming effects show effects of are shown in Table 2. Casually articulated (i.e., flapped) items
specificity. The observed effects of specificity in Experiment 3 were responded to more quickly than carefully articulated items,
may have arisen because lexical decision responses in the second F1(1, 46) ϭ 81.45, MSE ϭ 10,385.18, F2(1, 23) ϭ 53.82,
block tapped the recognition process before establishment of res- MSE ϭ 8,923.44. There was also a main effect of prime, F1(2,
onance between underlying and surface forms, not because the 92) ϭ 4.64, MSE ϭ 9,978.85, F2(2, 46) ϭ 3.61, MSE ϭ 7,867.89.
resonances were never established at all. Prime type and target type did not interact, F1 and F2 Ͻ 1.
To further evaluate the hypothesis that underlying representa- Planned comparisons based on the main effect of prime type
tions always resonate with surface forms, despite the fact that such revealed significant differences between the match and control
resonances may take time to develop, we conducted two further conditions and between the mismatch and control conditions, F1(1,
experiments in which we combined the shadowing and lexical 92) ϭ 4.31, F2(1, 46) ϭ 3.47, p ϭ .069 and F1(1, 92) ϭ 8.82, F2(1,
decision tasks. In Experiment 5, participants performed the easy- 46) ϭ 6.80, respectively. However, the difference between the
discrimination lexical decision task in the first block and the match and mismatch conditions was not significant, F1 and
shadowing task in the second block. In Experiment 6, the tasks F2 Ͻ 1.
were reversed (shadowing followed by easy-discrimination lexical
decision). Discussion
Recall that we observed specificity effects in Experiment 3, in
Overall, matched and mismatched primes produced significant
which we presented the easy-discrimination lexical decision task
facilitative effects on target RTs. These results replicated those for
in both the first and second blocks. If resonances between under-
the alveolar items in Experiments 1 and 2. Furthermore, these
lying and surface forms fail to develop in this task, we should
results confirm our earlier assumption that although we are able to
observe only specificity effects in long-term priming, regardless of
tap into the system at a point before the development of resonances
the task employed in the second block. Very simply, if underlying
between underlying and restored surface forms, processing con-
forms are not contacted in the first block, we would not expect
tinues and these resonances are eventually established.
priming for stimuli mismatching in articulation.
One final question now arises: Can we prime resonances be-
However, if underlying forms establish resonances in the first
tween input and underlying chunks to cause them to develop more
block even in the easy-discrimination lexical decision task, use of
the shadowing task in the second block should reveal activation of
underlying forms, given that we have already established that the 4
To investigate in more detail the trend toward specificity for the careful
shadowing task affords the opportunity for the underlying repre- items in Experiments 1, 2, and 4, we performed contrasts based on the
sentations to resonate with the surface forms. In Experiment 6, we nonsignificant Prime Type ϫ Target Type interactions. Despite the numer-
reversed the tasks, presenting the shadowing task followed by the ical trends in all three experiments, there was no statistical support for the
easy-discrimination lexical decision task. In this experiment, we conclusion that careful items resulted in more specificity than casual items.
5
asked whether effects of underlying representations can be found Because our argument rests on the hypothesis that depth of processing
with a task that typically taps into the system before the establish- mediates magnitude of specificity in long-term priming, we conducted
ment of the required resonances (the easy-discrimination lexical additional analyses directly comparing the results from Experiments 1– 4.
decision task). Specifically, we conducted a one-way ANOVA on the magnitude of
specificity (MOS; see Table 2) for the alveolar stimuli in Experiments 1– 4.
The main effect of experiment was significant. As expected, planned
Experiment 5 contrasts revealed that MOS for Experiment 3 was significantly larger than
in Experiments 1, 2, and 4. Moreover, none of the differences in MOS
Method
between Experiments 1, 2, and 4 were significant. This analysis confirms
Participants. A different group of 48 participants were recruited from that significantly larger specificity effects were obtained only in Experi-
the University at Buffalo community. They were paid $5 or received partial ment 3, as expected.
12. 550 MCLENNAN, LUCE, AND CHARLES-LUCE
quickly? In operational terms, can we induce priming of underly- performing the easy-discrimination lexical decision task in both
ing representations in an easy-discrimination lexical decision task Blocks 1 and 2, as had been done in Experiment 3, Experiment 6
by having participants shadow stimuli in the prime block? Despite participants performed the shadowing task in Block 1.
the fact that in Experiment 3 we found that easy-discrimination As predicted, preceding a superficial processing task with one
lexical decision produces pronounced specificity effects, we pre- that encourages contact with underlying representations results in
dict that, even in this task, resonances between input and under- a lack of specificity effects.6 Simply put, resonances prime: Con-
lying chunks can be made to develop more quickly if they have tacting an underlying representation makes it easier to establish
recently been established in a deeper processing task (i.e., that same resonance at a slightly later time (see also Grossberg &
shadowing). Meyers, 2000, p. 738).
Experiment 6 General Discussion
This investigation began with a simple question: Are flaps
Method mapped onto their underlying phonemic counterparts during per-
Participants. A different group of 48 participants were recruited from ceptual processing? If affirmative, the answer provides evidence
the University at Buffalo community. They were paid $5 or received partial against direct access models of recognition, instead supporting, in
credit for a course requirement. Participants met the same criteria as those part, the more traditional mediated models of speech perception
in Experiment 1. and spoken word recognition according to which the recoded
Materials. The materials in Block 1 were the same as those in Exper- speech waveform is mapped onto more abstract, underlying
iments 1 and 2, and the materials in Block 2 were the same as those in
representations.
Experiment 3.
The six long-term repetition priming experiments reported here
Procedure. The procedure was the same as in Experiment 5 except that
the order of tasks was reversed. In the first block participants performed a provide some evidence for mediated models (broadly construed;
shadowing task, and in the second block participants performed a lexical see subsequent discussion) while also suggesting the precise cir-
decision task. cumstances under which underlying representations may be con-
tacted during recognition. In Experiments 1 and 2, in which
Results participants shadowed flapped and carefully produced alveolar
stimuli in both the prime and target blocks, flaps primed carefully
RTs less than 500 ms or greater than 2,500 ms were replaced articulated stimuli and vice versa, a result consistent with the
with the appropriate condition mean. Less than 3% of the RTs notion that flaps activate their underlying phonemic counterparts.
were excluded from the analyses. In addition, 1 participant was Crucially, however, non-alveolar stimuli produced marked speci-
excluded. Accuracy was greater than 94% and produced no sig- ficity effects, suggesting that the long-term repetition priming
nificant outcomes. effect is not lemma based. Moreover, the finding that non-alveolar
RTs as a function of prime and target type are plotted in the stimuli prime only when they match on articulation style, whereas
lower right panel of Figure 4. Magnitudes of specificity and alveolar stimuli need not match to produce facilitative repetition
priming are shown in Table 2. Casually articulated (i.e., flapped) effects, suggests that the phonological and lexical ambiguity in-
items were responded to more quickly than carefully articulated herent in flapped stimuli is a necessary condition for activation of
items, F1(1, 46) ϭ 24.82, MSE ϭ 26,438.17, F2(1, 23) ϭ 20.71, underlying representations.
MSE ϭ 15,106.97. There was also a main effect of prime by
participants, F1(2, 92) ϭ 3.85, MSE ϭ 27,195.83, although the
6
effect failed to reach significance by items, F2(2, 46) ϭ 1.94, To compare more directly the effects of different study and test tasks
MSE ϭ 18,146.59, p ϭ .16. Prime type and target type did not (i.e., shadowing vs. lexical decision) on priming, we conducted a series of
comparisons across experiments in which we held test task constant while
interact, F1 and F2 Ͻ 1.
varying study task, and vice versa. For the comparisons in which test task
Planned comparisons based on the main effect of prime type
was held constant, RTs in Experiment 5 (lexical decision followed by
revealed significant differences by participants between the match shadowing) were compared with target RTs in Experiments 1 and 2
and control conditions and between the mismatch and control (shadowing followed by shadowing). In addition, target RTs in Experi-
conditions; the effects by items were statistically somewhat ment 6 (shadowing followed by lexical decision) were compared with
weaker, F1(1, 92) ϭ 5.42, F2(1, 46) ϭ 3.78, p ϭ .058 and F1(1, target RTs in Experiment 3 (lexical decision followed by lexical decision).
92) ϭ 6.12, F2(1, 46) ϭ 1.53, p ϭ .223, respectively. However, the The Prime Type ϫ Experiment interaction was not significant in the
difference between the match and mismatch conditions was not comparison of Experiments 1 and 2 with Experiment 5, indicating an
significant, F1 and F2 Ͻ 1. equivalent lack of specificity across these experiments. However, as ex-
pected, the Prime Type ϫ Experiment interaction was significant in the
comparison between Experiments 3 and 6, confirming that shadowing
Discussion during study attenuates specificity effects when participants make lexical
decisions during test. For the analyses in which study task was held
Overall, matched and mismatched primes produced facilitative
constant, the comparison of Experiments 1 and 2 with Experiment 6
effects on target RTs (although statistical support was somewhat
resulted in a nonsignificant Prime Type ϫ Experiment interaction, indi-
weaker than in previous experiments). These results replicated cating an equivalent lack of specificity across these experiments. However,
those for the alveolar items in Experiments 1 and 2 but differed as expected, the Prime Type ϫ Experiment interaction was significant in
from the results of Experiment 3 with the same stimuli and task in the comparison between Experiments 3 and 5, confirming that shadowing
Block 2. Indeed, the only difference between Experiments 3 and 6 attenuates specificity, even when participants are presented with the
was the task performed by participants in Block 1. Rather than specificity-inducing easy-lexical-discrimination task during study.
13. REPRESENTATION OF LEXICAL FORM 551
In Experiments 3 and 4, participants made lexical decisions in when flaps are present), when items are processed to a deep level
both the prime and target blocks to the same alveolar stimuli used (as in the shadowing task), and when enough time is allowed for
in Experiments 1 and 2. By manipulating the difficulty of word– the underlying representations to have an effect on recognition (as
nonword discrimination across the two experiments, we tested the in the hard-discrimination lexical decision task). Alternatively,
hypothesis that “depth” of processing may be crucial in activating surface representations appear to dominate processing when spo-
underlying representations. In Experiment 3, in which word– ken input is unambiguous (i.e., when non-alveolar stimuli are
nonword discrimination was made easier by the inclusion of un- used), when items are not processed to a deep level (as in easy-
wordlike nonwords, there was no evidence that flapped stimuli discrimination lexical decision), and when there is insufficient
activated their underlying phonemic counterparts: Flapped words time for the underlying representations to have an effect on rec-
failed to prime carefully articulated words, and vice versa. How- ognition (again as in the easy-discrimination lexical decision).
ever, in Experiment 4, in which discrimination was made difficult How, then, do we account for the activation of underlying forms
by the inclusion of nonwords that were very wordlike, clear evi- and the circumstances under which their effects are manifested in
dence for activation of underlying representations reemerged: the recognition process? We propose an account of these findings
Flaps primed carefully articulated words, and vice versa. Experi- based on Grossberg’s ARTPHONE model (Grossberg et al., 1997;
ments 3 and 4 suggest that depth of processing (manipulated see also Vitevitch & Luce, 1999). To review, acoustic–phonetic
through ease of lexical discrimination) mediates the activation of input composed of relatively veridical surface representations res-
underlying representations. onates with chunks corresponding to more abstract phonological
Finally, in Experiments 5 and 6, we crossed the shadowing and representations, as well as chunks corresponding to less abstract,
lexical decision tasks, using lexical decision in the prime block and allophonic representations. These resonances serve as the basis for
shadowing in the target block in Experiment 5 and the reverse long-term repetition priming.
arrangement of tasks in Experiment 6. The particular version of the In the absence of ambiguity in the input, the resonances between
lexical decision task used in these two experiments was the same surface forms and chunks corresponding to underlying represen-
one employed in Experiment 3 (i.e., the easy-discrimination task) tations preserve detail (see Grossberg & Meyers, 2000). However,
in which we observed no evidence of activation of underlying underlying representations (or chunks) activated by ambiguous
representations. Thus, we combined a task that consistently pro- flaps result in a restoration of surface representations not actually
duced activation of underlying forms (i.e., shadowing) with a task included in the input.
that produced no evidence of underlying activation (i.e., easy- Deep phonological processing associated with shadowing spo-
discrimination lexical decision). ken stimuli and making difficult word–nonword discriminations
In Experiment 5, in which the easy-discrimination lexical task encourages the restoration of surface representations by underlying
occurred in the prime block and the shadowing task occurred in the representations, and instantiation or restoration of the surface form
target block, we obtained evidence for activation of underlying by the underlying chunks requires time. Thus, tasks that tap into
representations. Thus, despite the fact that easy-discrimination the recognition process before restoration of the underlying form
lexical decision failed to produce activation of underlying forms in show strong effects of specificity in long-term repetition priming,
Experiment 3, stimuli in this task still acted as effective primes for presumably because the underlying representations may not have
underlying representations when the shadowing task was used in had sufficient time to establish resonance with a restored surface
the target block. This result suggests that underlying representa- form. However, the present evidence suggests that even though
tions are indeed contacted in the easy-discrimination lexical task: recognition may occur before establishment of resonances with
hence the priming effect. However, these underlying forms may underlying phonemic representations, these representations are
have little or no effect on processing when a more superficial nonetheless contacted, probably obligatorily. Moreover, it may be
analysis of the stimulus suffices (as in easy lexical discrimination possible to prime the resonances themselves, such that previous
in the target block of Experiment 3). activation of underlying forms makes establishing resonances with
In Experiment 6, in which the situation was reversed (shadow- these forms easier at a later time.
ing was used in the prime block and easy-discrimination lexical As an aside, we should note that what we have referred to
decision in the target block), we again obtained evidence (albeit throughout as depth of processing may or may not be coextensive
statistically somewhat weaker) for activation of underlying forms. with the time course of processing (see also Luce, McLennan, &
Note that this latter finding contrasts with Experiment 3, in which Charles-Luce, in press). It is not always the case that the fastest
easy-discrimination lexical decision in the target block produced responses result in the most specificity, although there is certainly
only specificity effects. Apparently, once the underlying forms a trend in that direction, especially within a given task. For
have been activated in the prime block, their effects are sufficiently example, compare the data for two lexical decision experiments in
strong and long-lasting to manifest themselves even in a task that which the faster responses resulted in marked specificity. How-
requires only superficial stimulus processing. ever, shadowing produced roughly equivalent RTs (in Experi-
Overall, we obtained a data pattern consistent with activation of ments 1, 2, and 5) to lexical decision (in Experiment 3) for the
the underlying phonemic counterparts of flaps during spoken word alveolar stimuli, yet shadowing consistently resulted in diminished
processing. Indeed, in only two circumstances did we observe specificity effects (relative to Experiment 3). At present, we can
specificity effects: (a) for the non-alveolar stimuli in Experi- only acknowledge that depth of processing may be strongly asso-
ments 1 and 2 and (b) for the alveolar stimuli when the easy- ciated with the time course of processing but may also encompass
discrimination lexical decision task was used in both prime and other variables, such as the need to produce a response.
target blocks. Underlying representations appear to dominate pro- The present results bear a marked resemblance to recent work
cessing when spoken input is phonologically ambiguous (i.e., reported by Halle, Chereau, and Segui (2000), who examined the
´ ´
14. 552 MCLENNAN, LUCE, AND CHARLES-LUCE
effects of voice assimilation in French on the perception of under- expectation and sensory input in which the percept may reside
lying phonemic forms. Because of voice assimilation in French, neither in the sensory data nor in the long-term representation but
voiced stops followed by voiceless segments, as in words such as in some melange of the two.
´
/absyrd/, are devoiced, as in [apsyrd]. Halle et al. examined
´
whether French participants would perceive the first consonant as
References
a /p/, which is actually present in the signal, or as a /b/, which is
consistent with both the underlying representation and the orthog- Bloomfield, L. (1933). Language. New York: Holt.
raphy. Using a phonemic gating task, they found that /p/ initially Chomsky, N., & Halle, M. (1968). The sound pattern of English. New
dominated participants’ responses, with /b/ responses gradually York: Harper & Row.
increasing over time and eventually overcoming the initial /p/ Church, B. A., & Schacter, D. L. (1994). Perceptual specificity of auditory
responses. In terms of the resonance framework, information priming: Implicit memory for voice intonation and fundamental fre-
present in the signal (in this case /p/) dominates processing until quency. Journal of Experimental Psychology: Learning, Memory, and
sufficient time has elapsed for the underlying /b/ to instantiate a Cognition, 18, 43–57.
Coltheart, M., Davelaar, E., Jonasson, J. T., & Besner, D. (1977). Access
corresponding surface representation. After the chunk correspond-
to the internal lexicon. In S. Dornic (Ed.), Attention and performance VI
ing to the underlying /b/ instantiates the appropriate surface rep- (pp. 535-555). Hillsdale, NJ: Erlbaum.
resentation (as a result of the learned association between [apsyrd] Fisher, C., Hunt, C., Chambers, K., & Church, B. (2001). Abstraction and
and underlying /b/), /b/ responses dominate. Although the authors specificity in preschoolers’ representations of novel spoken words. Jour-
were unable to determine definitively whether their effect had an nal of Memory and Language, 45, 665– 687.
orthographic or morphophonemic locus, their results are clearly Goldinger, S. D. (1996). Words and voices: Episodic traces in spoken word
consistent with the framework we are proposing. identification and recognition memory. Journal of Experimental Psy-
The present results may pose some difficulties, although pre- chology: Learning, Memory, and Cognition, 22, 1166 –1183.
sumably not insurmountable difficulties, to current connectionist Goldinger, S. D. (1998). Echoes of echoes? An episodic theory of lexical
models of spoken word recognition. TRACE and Shortlist, for access. Psychological Review, 105, 251–279.
Grossberg, S. (1986). The adaptive self-organization of serial order in
example, both lack an allophonic layer of representation, a mini-
behavior: Speech, language, and motor control. In E. C. Schwab & H. C.
mal requirement dictated by the finding that, under appropriate Nusbaum (Eds.), Pattern recognition by humans and machines: Vol. 1.
circumstances, flaps activate their phonemic counterparts. Only Speech perception (pp. 187–294). New York: Academic Press.
PARSYN incorporates an explicit allophonic level. However, Grossberg, S. (1995). The attentive brain. American Scientist, 83, 438 –
PARSYN lacks phonemic representations, which may prove prob- 449.
lematic in accounting for the activation of underlying forms (al- Grossberg, S. (1999). The link between attention, brain learning, and
though PARSYN’s lexical representations are phonemically consciousness. Consciousness and Cognition, 8, 1– 44.
coded). Grossberg, S., Boardman, I., & Cohen, M. (1997). Neural dynamics of
Although in their current forms, TRACE, Shortlist, and variable-rate speech categorization. Journal of Experimental Psychol-
PARSYN all may have some difficulties in accounting for the ogy: Human Perception and Performance, 23, 483–503.
Grossberg, S., & Meyers, C. W. (2000). The resonant dynamics of speech
complete set of results, nothing in their architectures prohibits the
perception: Interword integration and duration-dependent backward ef-
necessary modifications: TRACE and Shortlist could add an allo- fects. Psychological Review, 107, 735–767.
phonic level, and PARSYN could add a phonemic level. However, Grossberg, S., & Stone, G. O. (1986). Neural dynamics of word recognition
even with the appropriate representations, it is unclear how these and recall: Attentional priming, learning, and resonance. Psychological
or similar models could account for the observed effects of depth Review, 93, 46 –74.
of processing. One possibility may be to incorporate an attentional Halle, P. A., Chereau, C., & Segui, J. (2000). Where is the /b/ in “absurde”
´ ´
focus by manipulating weights at various levels. Overall, however, [apsyrd]? It is in French listeners’ minds. Journal of Memory and
we believe that the adaptive resonance framework most naturally Language, 43, 618 – 639.
handles the range of observed effects in the present studies. Harris, Z. S. (1955). From phoneme to morpheme. Language, 31, 190 –
We now come full circle to ask how the present results bear on 222.
Jackson, A., & Morton, J. (1984). Facilitation of auditory word recognition.
the distinction between mediated and direct access models, the
Memory & Cognition, 12, 568 –574.
original theoretical focus of our work. Clearly, one aspect of Jusczyk, P., & Luce, P. A. (2002). Speech perception and spoken word
mediated models has been supported, namely the activation of recognition: Past and present. Ear & Hearing, 23, 1– 40.
underlying abstract forms in spoken word processing. However, Kempley, S. T., & Morton, J. (1982). The effects of priming with regularly
the adaptive resonance framework we have adopted bears a strong and irregularly related words in auditory word recognition. British
resemblance to a direct access model. After all, we have proposed Journal of Psychology, 73, 441– 454.
that veridical representations first make contact with fairly specific Kenstowicz, M., & Kisseberth, C. (1979). Generative phonology: Descrip-
(i.e., allophonic) representations, only after which do underlying tion and theory. New York: Academic Press.
forms come into play. Moreover, the results for the non-alveolar Klatt, D. H. (1989). Review of selected models of speech perception. In W.
stimuli suggest that fairly specific representations dominate pro- Marslen-Wilson (Ed.), Lexical representation and process (pp. 169 –
226). London: MIT Press.
cessing in the absence of phonological ambiguity or deeper pro-
Kucera, H., & Francis, W. (1967). Computational analysis of present day
ˇ
cessing. In short, the adaptive resonance framework requires us to American English. Providence, RI: Brown University Press.
reconceptualize the problem. Indeed, neither of the prototypical Lahiri, A., & Marslen-Wilson, W. (1991). The mental representation of
models illustrated in Figure 1 can adequately account for the lexical form: A phonological approach to the recognition lexicon. Cog-
present data, in part because they fail to acknowledge that percep- nition, 38, 245–294.
tion may be better characterized as a resonance between learned Luce, P. A., Goldinger, S. D., Auer, E. T., & Vitevitch, M. S. (2000).